Europ . J . Combinatorics (1996) 17 , 209 – 221 On the Complexity of Recognizing Hamming Graphs and Related Classes of Graphs W ILFRIED I MRICH AND S ANDI K LAVZ ä AR This paper contains a new algorithm that recognizes whether a given graph G is a Hamming graph , i . e . a Cartesian product of complete graphs , in O ( m ) time and O ( n 2 ) space . Here m and n denote the numbers of edges and vertices of G , respectively . Previously this was only possible in O ( m log n ) time . Moreover , we present a survey of other recognition algorithms for Hamming graphs , retracts of Hamming graphs and isometric subgraphs of Hamming graphs . Special emphasis is also given to the bipartite case in which these classes are reduced to binary Hamming graphs , median graphs and partial binary Hamming graphs . ÷ 1996 Academic Press Limited 1 . I NTRODUCTION This paper is a contribution to the recognition of classes of graphs defined by metric properties . These classes include Hamming graphs , quasi-median graphs , partial Hamming graphs , binary Hamming graphs , median graphs and partial binary Hamming graphs . We shall define the above-mentioned classes of graphs , list some of their structural properties , in particular those which are exploited from the algorithmic point of view , and then shortly describe the ideas behind several recognition algorithms . Moreover , we also present a new recognition algorithm which is optimal in its time complexity for the recognition of Hamming graphs . All graphs considered in this paper are finite undirected graphs without loops or multiple edges . Throughout the paper , for a given graph G , let n and m denote the number of its vertices and edges , respectively . For a graph G and a vertex set X ’ V ( G ) , let k X l denote the subgraph of G induced by X . A subgraph H of a graph G is an isometric subgraph if d H ( u , y ) 5 d G ( u , y ) for all u , y P V ( H ) . In addition , if a : V ( H ) 5 V ( G ) maps edges into edges and if a ( H ) is an isometric subgraph of G , we call a an isometric embedding of H into G . The intery al I ( u , y ) between vertices u and y consists of all vertices on shortest paths between u and y . A subgraph H of G is cony ex if , for any u , y P V ( H ) , I ( u , y ) ‘ V ( H ) . Clearly , a convex subgraph is an isometric subgraph , but the converse need not be true . The cony ex hull of a subgraph H in G is the intersection of all convex subgraphs of G containing H , i . e . the smallest convex subgraph containing H . The Cartesian product G H of graphs G and H is the graph with vertex set V ( G ) 3 V ( H ) and ( a , x )( b , y ) P E ( G H ) whenever ab P E ( G ) and x 5 y , or a 5 b and xy P E ( H ) . A mapping f : V ( G ) 5 V ( H ) is a graph homomorphism if f ( u ) f ( y ) P E ( H ) whenever u y P E ( G ) . A subgraph H of a graph G is a retract of G if there is a homomorphism r from V ( G ) to V ( H ) such that r ( y ) 5 y for every y P V ( H ) . The map r is called a retraction . If we allow that r maps an edge of G either to an edge or to a single vertex in H , we call H a weak retract of G and r a weak retraction . 209 0195-6698 / 96 / 020209 1 13 $18 . 00 / 0 ÷ 1996 Academic Press Limited 210 W . Imrich and S . Klay zä ar 2 . T HE B IPARTITE OR B INARY C ASE In this section we consider binary Hamming graphs , their retracts and isometric subgraphs . 2 . 1 . Binary Hamming graphs . Binary Hamming graphs are also known as hyper- cubes . A d -dimensional hypercube Q d ( d - cube for short) is a graph the vertices of which are all d -tuples b 1 b 2 ? ? ? b d with b i P h 0 , 1 j and two vertices are adjacent if the corresponding tuples dif fer in precisely one co-ordinate . Alternatively , Q d is the Cartesian product of d copies of K 2 . Clearly , Q d is a connected , bipartite d -regular graph on n 5 2 d vertices and m 5 d ? 2 d 2 1 edges . Moreover , its automorphism group is vertex and edge transitive . We also note that the usual shortest path distance between any two vertices x and y of Q d is the number of positions in which x and y dif fer . For example , the distance between 0110 and 1101 is 3 . This distance is also called the Hamming distance between x and y . Bhat [9] proposed an O ( m ) algorithm for recognizing binary Hamming graphs . In fact , his algorithm can essentially be obtained by specializing the algorithm for Hamming graphs which we will present in Section 3 . 1 . Here we propose an alternative algorithm of the same time complexity for recognizing binary Hamming graphs . One can obtain it by applying an algorithm from [26] or an algorithm from [22] for recognizing median graphs . The time complexities of these two algorithms , which will be discussed in the next section , can be reduced for binary Hamming graphs because we need not worry about convexity . Let u y be an arbitrary edge of Q d . To fix ideas , let u 5 00 ? ? ? 0 and y 5 10 ? ? ? 0 . Then the vertices the first co-ordinate of which is zero are exactly those vertices V u y i n Q d which are closer to u than to y . Furthermore , they induce a ( d 2 1)-cube k V u y l i n Q d . Analogously , we obtain a ( d 2 1)-cube k V y u l on the set V y u of vertices that are closer to y than to u , i . e . those vertices the first co-ordinate of which is 1 . The edges not in k V u y l or k V y u l are of the form (0 x 2 x 3 ? ? ? x d )(1 x 2 x 3 ? ? ? x d ) . It is easily seen that these edges are a matching of Q d and that this matching defines an isomorphism a : 0 x 2 x 3 ? ? ? x d S 1 x 2 x 3 ? ? ? x d of k V u y l onto k V y u l . This information already suf fices for an O ( n log n ) algorithm for recognizing binary Hamming graphs . For , let G be a connected graph on n vertices . If it is a binary Hamming graph its number m of edges must be ( n / 2) log 2 n 5 O ( n log n ) . First , we can check in that many steps whether G is bipartite . Then , choosing u y P E ( G ) arbitrarily , we can obtain all distances d G ( u , x ) and d G ( y , x ) for x P V ( G ) in 2 m steps . Thus , V u y and V y u can be determined in O ( m ) time . In another O ( m ) steps one can determine whether the edges not in k V u y l or k V y u l define a matching and an isomorphism between k V u y l and k V y u l . It remains to show that k V u y l which has less than m / 2 edges is a binary Hamming graph . The complexity thus is O ( m ) 1 O ( m / 2) 1 O ( m / 4) 1 ? ? ? 1 O (1) 5 O ( m ) . Since all edges have to be checked , this complexity is best possible . We formulate this as a theorem . T HEOREM 2 . 1 [9] . For a giy en graph G on n y ertices , one can decide in O ( n log n ) steps whether G is a binary Hamming graph . This complexity is optimal . Recognizing Hamming graphs 211 For the next two sections we recall that for any subgraph of a binary Hamming 1– graph , m < 2 n log n , cf . [1 , 18] . 2 . 2 . Median graphs . A median of a set of three vertices , u , y and w , is a vertex that lies in I ( u , y ) > I ( u , w ) > I ( y , w ) . In other words , x is a median of u , y and w if d ( u , x ) 1 d ( x , y ) 5 d ( u , y ) , d ( y , x ) 1 d ( x , w ) 5 d ( y , w ) , d ( u , x ) 1 d ( x , w ) 5 d ( u , w ) . Let u y P E ( G ) and let w P V ( G ) . It is easy to verify that u , y and w have a median if f d ( u , w ) ? d ( y , w ) . This observation in turn implies that G is bipartite if any three vertices of a graph G have a median . A connected graph G is a median graph if every triple of its vertices has a unique median . It is easily seen that binary Hamming graphs and trees are median graphs . Furthermore , Bandelt proved the following theorem . T HEOREM 2 . 2 [6] . A graph is a median graph if f it is a retract of a binary Hamming graphs . In fact , one could also replace retracts by weak retracts in Theorem 2 . 2 . We also note that a retract always is an isometric subgraph . Next , we describe the convex expansion procedure due to Mulder [27 , 28] which leads to another characterization of median graphs .